Summary of three-dimensional point cloud target extraction

three-dimensional point cloud target extraction (cont.) 3. Three-dimensional point cloud target extraction 3.1 General FlowSummarize the general steps of goal extraction based on your personal knowledge:
As shown above, the three-dimensional point cloud target extraction of the key two steps is: feature extraction and selection, classification, is not the entire method process and the image of the target recognition a bit like. Essentially, the method flow is basically the same for any target recognition involved. Why do feature extraction, because we want to identify the goal is generally in a large scene, the various goals are mixed, since to identify a target, of course, there is a need to have an indicator or numerical value to maximize the difference between the different goals, the indicator or the value is the so-called target characteristics. Therefore, when we identify the target, we often have to adopt the characteristics that are suitable for this goal. Like the convolutional neural network in image recognition. CNN, why it is compared with the traditional hand-designed features to identify the recognition rate is high, the essential reason is that CNN is characterized by learning, and the characteristics of the classifier is combined with the optimization. Classifiers are not cumbersome, SVM, boosting, decision trees, and so on. 3.2 Feature ExtractionThe importance of feature extraction from the above can be seen, it is the final result can not meet the expectations of the most important factor. To see how the different classifiers affect the results:
The above classifier is the nearest neighbor, decision Tree, two discriminant analysis, SVM, the characteristics used are the same. From the results, it can be concluded that the decisive factor affecting the target recognition is the feature extraction. In fact, that's why I understand why CNN uses the Softmax classifier, which is not a classifier, but a feature that affects the results. Well, let's summarize the features of the three-dimensional point cloud. <1>2d features

2d Geometric features: radius, point density, area ...

2d Local shape Features

Based on the features of the cumulative graph: Grid dot number, elevation difference, elevation standard deviation ... <2>3d features

3d Geometric features: radius, elevation difference, elevation standard deviation, point density

3d Local shape Features: Linear features, planar features, scattered features, total variance, anisotropy, feature entropy, eigenvalues and curvature ...

<3> Texture Features

RGB Strength <4> chart features

Point feature histogram PFH, fast Point feature histogram FPFH, viewpoint feature histogram vfh,signature

of histograms of Orientations (SHOT) <5> others

Spin Images,global fpfh,global radius-based Surface Descriptor (GRSD), globalstructure histogram (GSH)

3.3 classifiers

Classifier this or omit it, only with the enthusiasm of everyone, the result is not improved, the following are some commonly used classifiers:

4. Summary

Three-dimensional point cloud target extraction personally think there is still a lot of research prospects, after all, the overall target recognition rate is not very high, there are many areas that need improvement. As previously mentioned, the effect of the entire result is the feature extraction of this step, the individual feel can from this step to carry out research, I have done some related experiments, is combined with deep learning, (there is time to introduce this piece) see some foreign mainstream periodicals, seemingly do not many people. Also, personally think it is best to combine the three-dimensional point cloud with the data from other sensors to study, especially the image, the value of research is definitely the leverage. If you want to know some specific research results, we can go to mainstream journals to see, of course, you can ask me, or need any information can also, for the country to contribute ~ ~ ~